Add FakeNews Gradio interface

README.md

@@ -1,13 +1,21 @@
 ---
-title: FakeNews Interface
-emoji: 🏢
-colorFrom: yellow
-colorTo: yellow
+title: FakeNews Classifier
 sdk: gradio
 sdk_version: 6.19.0
-python_version: '3.13'
 app_file: app.py
-pinned: false
+license: mit
+models:
+  - JayNightmare/FakeNews
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# FakeNews Classifier
+
+Gradio interface for the `JayNightmare/FakeNews` PEFT LoRA adapter.
+
+The model repo stores adapter files in `adapter/`, so the Space loads the base model from the adapter config and attaches the adapter with `subfolder="adapter"`.
+
+Optional Space environment variables:
+
+- `MODEL_REPO`: defaults to `JayNightmare/FakeNews`
+- `ADAPTER_SUBFOLDER`: defaults to `adapter`
+- `BASE_MODEL_ID`: optional override if the adapter config points to a local path

app.py

@@ -0,0 +1,219 @@
+import json
+import os
+import threading
+
+import gradio as gr
+import torch
+from huggingface_hub import hf_hub_download
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+
+MODEL_REPO = os.getenv("MODEL_REPO", "JayNightmare/FakeNews")
+BASE_MODEL_ID = os.getenv("BASE_MODEL_ID")
+ADAPTER_SUBFOLDER = os.getenv("ADAPTER_SUBFOLDER", "adapter")
+
+ADAPTER_SUBFOLDER_CANDIDATES = tuple(
+    dict.fromkeys(
+        [
+            ADAPTER_SUBFOLDER,
+            "adapter",
+            "FakeNews_Model/adapter",
+            "artifacts/FakeNews_Model/adapter",
+            "model_adapter",
+            "FakeNews_Model/model_adapter",
+            "artifacts/FakeNews_Model/model_adapter",
+            "",
+        ]
+    )
+)
+
+generation_lock = threading.Lock()
+
+
+def download_adapter_config():
+    last_error = None
+
+    for subfolder in ADAPTER_SUBFOLDER_CANDIDATES:
+        try:
+            path = hf_hub_download(
+                repo_id=MODEL_REPO,
+                filename="adapter_config.json",
+                subfolder=subfolder or None,
+            )
+            return path, subfolder or None
+        except Exception as error:
+            last_error = error
+
+    tried = ", ".join(
+        subfolder or "<repo root>" for subfolder in ADAPTER_SUBFOLDER_CANDIDATES
+    )
+    raise RuntimeError(
+        "Could not find adapter_config.json. Set ADAPTER_SUBFOLDER to the repo "
+        "folder that contains adapter_config.json and adapter_model.safetensors. "
+        f"Tried: {tried}."
+    ) from last_error
+
+
+def resolve_base_model_id(adapter_config_path):
+    with open(adapter_config_path, "r", encoding="utf-8") as file:
+        adapter_config = json.load(file)
+
+    base_model_id = BASE_MODEL_ID or adapter_config.get("base_model_name_or_path")
+
+    if not base_model_id:
+        raise RuntimeError(
+            "Could not resolve the base model. Set BASE_MODEL_ID in the Space environment."
+        )
+
+    if base_model_id.startswith(("/", "./", "../")) and not BASE_MODEL_ID:
+        raise RuntimeError(
+            "adapter_config.json points to a local base model path. "
+            "Set BASE_MODEL_ID to the original Hugging Face base model id."
+        )
+
+    return base_model_id
+
+
+def load_tokenizer(base_model_id, adapter_subfolder):
+    try:
+        return AutoTokenizer.from_pretrained(
+            MODEL_REPO,
+            subfolder=adapter_subfolder,
+            trust_remote_code=True,
+        )
+    except Exception:
+        return AutoTokenizer.from_pretrained(
+            base_model_id,
+            trust_remote_code=True,
+        )
+
+
+def load_model():
+    adapter_config_path, adapter_subfolder = download_adapter_config()
+    base_model_id = resolve_base_model_id(adapter_config_path)
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    torch_dtype = torch.float16 if device == "cuda" else torch.float32
+
+    tokenizer = load_tokenizer(base_model_id, adapter_subfolder)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    base_model = AutoModelForCausalLM.from_pretrained(
+        base_model_id,
+        torch_dtype=torch_dtype,
+        trust_remote_code=True,
+    )
+
+    model = PeftModel.from_pretrained(
+        base_model,
+        MODEL_REPO,
+        subfolder=adapter_subfolder,
+    )
+    model.to(device)
+    model.eval()
+
+    return model, tokenizer, device, base_model_id, adapter_subfolder
+
+
+model, tokenizer, device, base_model_id, adapter_subfolder = load_model()
+
+
+def build_prompt(claim, context):
+    context = context.strip() or "No additional context provided."
+    messages = [
+        {
+            "role": "system",
+            "content": (
+                "You are FakeNews, a misinformation classifier. "
+                "Classify the claim using the available context. "
+                "Return a concise label and explanation."
+            ),
+        },
+        {
+            "role": "user",
+            "content": (
+                f"Claim:\n{claim.strip()}\n\n"
+                f"Context:\n{context}\n\n"
+                "Classify this as fake, real, misleading, or uncertain."
+            ),
+        },
+    ]
+
+    try:
+        return tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True,
+        )
+    except Exception:
+        return (
+            "System: You are FakeNews, a misinformation classifier.\n\n"
+            f"User: Claim:\n{claim.strip()}\n\n"
+            f"Context:\n{context}\n\n"
+            "Classify this as fake, real, misleading, or uncertain.\n\n"
+            "Assistant:"
+        )
+
+
+def classify(claim, context):
+    if not claim or not claim.strip():
+        return "Enter a claim to classify."
+
+    prompt = build_prompt(claim, context or "")
+    inputs = tokenizer(
+        prompt,
+        return_tensors="pt",
+        truncation=True,
+        max_length=2048,
+    ).to(device)
+
+    with generation_lock:
+        with torch.no_grad():
+            output_ids = model.generate(
+                **inputs,
+                max_new_tokens=160,
+                do_sample=False,
+                pad_token_id=tokenizer.pad_token_id,
+                eos_token_id=tokenizer.eos_token_id,
+            )
+
+    generated_ids = output_ids[0][inputs["input_ids"].shape[-1] :]
+    return tokenizer.decode(generated_ids, skip_special_tokens=True).strip()
+
+
+description = (
+    f"Loaded adapter from `{MODEL_REPO}`"
+    + (f" subfolder `{adapter_subfolder}`" if adapter_subfolder else "")
+    + f" on base model `{base_model_id}`."
+)
+
+demo = gr.Interface(
+    fn=classify,
+    inputs=[
+        gr.Textbox(label="Claim", lines=3, placeholder="Enter a news claim..."),
+        gr.Textbox(
+            label="Optional context",
+            lines=6,
+            placeholder="Paste article or source context here...",
+        ),
+    ],
+    outputs=gr.Textbox(label="FakeNews result", lines=8),
+    title="FakeNews Classifier",
+    description=description,
+    examples=[
+        [
+            "The Eiffel Tower was originally built for the 1889 World's Fair.",
+            "",
+        ],
+        [
+            "A viral post claims drinking salt water cures all viral infections.",
+            "No credible medical source supports this claim.",
+        ],
+    ],
+)
+
+
+if __name__ == "__main__":
+    demo.queue().launch()

requirements.txt

@@ -0,0 +1,7 @@
+accelerate>=1.7.0
+gradio>=6.19.0
+huggingface_hub>=0.35.0
+peft>=0.16.0
+safetensors>=0.4.5
+torch>=2.7.1
+transformers>=4.52.4